model

The guessing machine itself. The file, not the website.

training

The very long, very expensive part where it learned to guess. Already over by the time you type anything.

inference

The part that happens when you actually use it. Cheap once, enormous in bulk.

token

A chunk of a word. It guesses these, not whole words.

hallucination

Made it up. Sounded fine. Same machinery as a correct answer.

context window

How much it can hold in its head at once. Past the edge, it forgets.

memory

What it learned in training, plus what you pasted into this chat. Not your inbox. Not today's database.

retrieval

Fetching text and handing it over to read. The model does not go do anything. It reads.

RAG

A common way to do retrieval automatically: search, then stuff the hits into the prompt.

tool use

The model calls a function that runs somewhere else and comes back with a result. Can be look-up. Can be a change.

MCP

The agreed shape of the plug for tools. Not the capability. The connector, so a tool built once can be reused.

tool call

The moment it stops talking and presses a button in something else.

provenance

Which bits were fetched, which were calculated, which were guessed. Same calm voice, different reliability.

agency

Letting it choose the next tool call for a while. Looking something up is not the same as doing it.

agent

The same machine, allowed to press buttons on its own for a while.

fine-tuning

Extra practice on one narrow kind of thing, after the main learning is done.

GPU

The kind of chip it runs on. Fast at doing one small sum a very large number of times.

data center

The building. Machines, power, and a lot of equipment whose only job is moving heat outside.

interconnection

Permission and wire to draw power from the grid. The slow part. Years, not months.